How to remove people from travel photos with AI, and what gets invented

You waited ten minutes for the square in front of the old church to clear, and it never did. Now the picture has a family in matching rain jackets on the left, a man with a selfie stick in the middle distance and a tour group's flag poking into the sky. The building is perfect. Here is how to get the people out of it, step by step, and where the result stops being your photo.

Step 1: decide which tool the photo needs

The catalog has two kinds of model for this job, and they behave very differently.

• An object removal model (an eraser). You give it the photo and a mask that marks what has to go. It fills only the marked area and leaves every other pixel as it was. Type "eraser" into the search on the models page to find it.
• A general edit model. You give it the photo and a sentence such as "remove the people". It understands the request without a mask, but it returns a new picture: the whole frame is redrawn, and small details outside the people can shift.

Rule of thumb: if the photo matters as a photo (a print for the wall, a picture you will look at closely), use the eraser. If you need a quick clean version for a slide or a social post and you do not want to draw anything, try the edit model first.

Step 2: make the mask (eraser route)

The eraser asks for two files in the studio: the photo and a mask image. The studio has no brush for drawing a mask, so you make it in any image editor you already have, even a free one on your phone:

1. 01Open the photo and add a new layer filled with black.
2. 02With a soft brush in white, paint over each person you want gone, slightly wider than their outline. Include their shadow on the ground and any reflection in a window or a puddle; if the shadow stays, the eye finds a ghost.
3. 03Hide the photo, export only the black and white layer, at exactly the photo's size.
4. 04Upload both files and run. If the first result erased the wrong part, invert the mask and run again.

Step 3: or write one clear instruction (edit route)

An edit model reads a sentence, so the usual rules of edit prompts apply: say what changes, say what stays, one change per round. The guide to writing edit prompts goes deeper; for this job the wording is short.

Before you run, open the settings. Some edit models have an input fidelity option (keep it high so the building stays faithful) and a size option, because an edit model returns the picture at its own sizes, which may not match your original. If the output is smaller than the photo you started with, that is the reason.

Step 4: check the seams before you share it

Look at the result at full size, not as a thumbnail. The common faults are easy to spot once you know them: a repeated pattern of paving stones where the model copied the same patch twice, a window frame that bends where it passes behind the erased person, a lamp post that ends in mid-air, a blurry smear where a crowd used to be. Re-run with a slightly different mask, or erase in two smaller passes instead of one big one. Each run shows its price on the Generate button first, so a second attempt is a known cost; the post on the cost per usable picture covers budgeting for retries.

Moving footage: the video eraser

For a short clip of the same square, there is a video object removal model that takes the video and a plain word or phrase naming what to remove, such as "person on the left" or "tour flag", with no mask at all. A setting decides whether the original sound is kept. Expect more re-rolls than with a still: the fill has to stay consistent from frame to frame, and a flicker where a person used to walk is the typical failure.

What the AI invents, and what it cannot do

Nothing in the photo shows what was behind the man with the selfie stick, so the model guesses. The paving, the door, the bottom half of a statue it draws are plausible, not real. That is fine for a holiday album or a background for your own poster, but it means the edited picture is no longer a record of the place: do not use it as evidence, for journalism, or in a listing that promises how a property looks. Large gaps are the hard case. A person standing close to the camera may hide a third of the frame, and no eraser can rebuild that convincingly; crop or choose another shot instead. Dense crowds tend to turn into smudges. Text on signs behind the person often comes back as nonsense letters, as described in the post on hands and text. And a removed person can leave traces the mask did not cover: a hand on a railing, a bag on a bench, a second reflection. Keep the original file next to the edited one so you always know which is which.